Muyan-TTS Make Podcasts with AI Model Locally Step-by-Step Tutorial

Muyan-TTS Make Podcasts with AI Model Locally Step-by-Step Tutorial

---

---

AI Summary

Overview

• Video by Fahad Miraza on Moan TTS, a text-to-speech model optimized for podcast scenarios.

Key Features

• Open-source and trainable model suitable for zero- and one-shot use cases.
• Two model versions: Base TTS model (multispeaker) and Supervised Fine-tuning (single speaker).
• Efficient for voice cloning with lightweight fine-tuning capabilities.

Installation Instructions

1. Environment Setup:
   • Install FFmpeg multimedia library.
   • Create a virtual environment using Python 3.10.
2. Clone the Repository:
   • Link for repository in video description.
   • Install PyAudio using the command: pip install pyaudio.
3. Directory Structure:
   • Create a directory for pre-trained models with specified subdirectories.
   • Download the base model, SFT model, and Chinese Hubert model.
   • Log into Hugging Face CLI: huggingface-cli login and fetch the access token.
4. Running the Model:
   • Use tts.py script for text-to-speech conversion.
   • Provide sample audio and corresponding text for voice synthesis.

Performance

• Generates audio at approximately 1 second of audio per 30 seconds on standard GPUs.
• VRAM consumption during inference noted below 8 GB under typical usage.

Conclusion

• Moan TTS shows promise as a viable model for podcasting and voice cloning, with efficiency in deployment and adaptability for user-specific voices.